We have developed a simulation environment called CreatureSpace that allows us to test our agent theories on intelligent agents in a complex realistic environment. We present the CreatureSpace architecture and our experiences in combining multiple artificial intelligence techniques in a uniform environment. CreatureSpace is a combination of a realistic environment and intelligent agents that populate this environment using Half-Life as a rendering engine. The agents we create are deliberative agents with the ability to exist in a realistic environment. They can process what they see and behave according to this and their own desires. We focus on how agents can manage large amounts of information and describe our embedded knowledge solution. We have tested the ideas in the paper by using CreatureSpace to run fire evacuation simulations with synthetic characters developed using our embedded knowledge solution. We find that combining embedded knowledge and sketchy planning results in simple solutions to potentially difficult problems in realistic environments.